SOM.py

    ##################################################
#                                                #
#  Copyright 2009 -Marcel Pinheiro Caraciolo-    #
#  Email: caraciol@gmail.com                     #
#                                                #
#  --The Self-Organizing-maps Neural networks.   #
#  -- Version: 0.1  - 21/04/2009                 #
##################################################

"""
     Self Organizing maps neural networks - SOM Sample example.
"""
import random
import math

class Map(object):
    
    #Class Constructor
    #@param dimensions: Number of input dimensions
    #@param length: Length of the output grid
    #@param filePath: The file path with the input data.
    def __init__(self,dimensions,length,filePath):
        #Collection of weights
        self.outputs = []
        #Current iteration
        self.iteration = 0
        #Side length of output grid.
        self.length = length
        #Number of input dimensions
        self.dimensions = dimensions
        #Random Generator
        self.rand = random.Random()
        #Label<Patterns> dict
        self.patterns = {}
        #Initialise the SOM
        self.initialise()
        #Load the dataset
        self.loadData(filePath)
        #Normalise the patterns
        self.normalisePatterns()
        #Training method
        self.train(0.0000001)
        #Dump the coordinates
        self.dumpCoordinates()
    
    #Initialise the Network
    def initialise(self):
        #10x10 Dimensional Grid
        for i in xrange(self.length):
            self.outputs.append([])
            for j in xrange(self.length):
                self.outputs[i].append(Neuron(i,j,self.length))
                for k in xrange(self.dimensions):
                    self.outputs[i][j].weights.append(self.rand.random()) 
    
    #Load the dataset
    #@param filePath: The file path
    def loadData(self,filePath):
        fileHandle = open(filePath)
        #Ignore the first line (header)
        fileList = fileHandle.readlines()[1:]
        for line in fileList:
            line = line.strip()
            lineSet = line.split(',')
	    print lineSet,'\n'
            self.patterns[lineSet[0]] = []
            for i in xrange(self.dimensions):
                self.patterns[lineSet[0]].append(float(lineSet[i+1]))
        fileHandle.close()

    #Normalise the patterns
    def normalisePatterns(self):
        for j in xrange(self.dimensions):
            sum = 0.0
            for key in self.patterns.keys():
                sum += self.patterns[key][j]
            average = sum / float(len(self.patterns))
            
            for key in self.patterns.keys():
                self.patterns[key][j] = self.patterns[key][j] / average
    
    #The training method
    #@param maxError: the error treshold
    def train(self,maxError):
        currentError = 100000.0
        
        while currentError > maxError:
            currentError = 0.0
            trainingSet = []
            for pattern in self.patterns.values():
                trainingSet.append(pattern)
            for i in xrange(len(self.patterns)):
                pattern = trainingSet[self.rand.randrange(len(self.patterns)-i)]
                currentError += self.trainPattern(pattern)
                trainingSet.remove(pattern)
            
            print "Current Error: %.7f" % (currentError,)
    
    #Train Pattern
    #@param pattern: The input pattern
    def trainPattern(self,pattern):
        error = 0.0
        winner = self.winner(pattern)
        for i in xrange(self.length):
            for j in xrange(self.length):
                error += self.outputs[i][j].updateWeights(pattern,winner,self.iteration)
        
        self.iteration+=1
        
        return abs(error / (self.length * self.length))
    
    #The winner takes all rule
    #@param pattern: the input pattern
    #@return the neuron winner
    def winner(self,pattern):
        winner = None
        minD = 10000000.0
        for i in xrange(self.length):
            for j in xrange(self.length):
                d = self.distance(pattern,self.outputs[i][j].weights)
                if d < minD:
                    minD = d
                    winner = self.outputs[i][j]
        return winner

    #The euclidian distance
    #@param inVector: the input Pattern
    #@param outVector: the output neurons vector
    def distance(self,inVector,outVector):
        value = 0.0
        for i in xrange(len(inVector)):
            value +=  pow((inVector[i] - outVector[i]),2)
        
        return math.sqrt(value)
    
    #Dump the coordinates
    def dumpCoordinates(self):
        for key in self.patterns.keys():
            n = self.winner(self.patterns[key])
            print "%s,%d,%d" % (key,n.X,n.Y)
 

#Simple SOM Neuron
class Neuron(object):
    #Class Constructor
    #@param x : X Coordinate
    #@param y : Y Coordinate
    #@param length: The Strength
    def __init__(self,x,y,length):
        #Neuron weights
        self.weights = []
        #X Coordinate
        self.X = x
        #Y Coordinate
        self.Y = y
        #Length 
        self.length = length
        #nf
        self.nf = 1000 / math.log(length)
    
    #Gaussian Equation
    #@param winner: The neuron winner
    #@param iteration: current iteration
    def gauss(self,winner,iteration):
        distance = math.sqrt(pow(winner.X - self.X,2) + pow(winner.Y - self.Y,2))
        
        return math.exp(-pow(distance,2) / (pow(self.strength(iteration),2)))
    
    #Set the learning rate
    #@param iteration: The current iteration 
    def learningRate(self,iteration):
        return math.exp(-iteration/1000) * 0.1
 
    #Set the strength (Neighborhood rate)
    #@param iteration : the current iteration
    def strength(self,iteration):
        return math.exp(-iteration/self.nf) * self.length
    
    #Update the weights
    #@param pattern: The input pattern
    #@param winner: The neuron winner
    #@param iteration: The current iteration
    def updateWeights(self,pattern,winner,iteration):
        sum = 0.0
        for i in xrange(len(self.weights)):
            delta = self.learningRate(iteration) * self.gauss(winner,iteration) * (pattern[i] - self.weights[i])
            self.weights[i] += delta
            sum += delta
        
        return sum / len(self.weights)

    

#####################################
#import psyco

#psyco.full()
map = Map(3,20,"dataset.txt")